Modeling, Evaluating, and Embodying Personality in LLMs: A Survey

Reviews, theory, and governance2025ACL AnthologyApproved editorial review

Authors: Iago Alves Brito, Julia Soares Dollis, Fernanda Bufon Färber, Pedro Schindler Freire Brasil Ribeiro, Rafael Teixeira Sousa, Arlindo Rodrigues Galvão Filho

Keywords: Large Language Models, Personality, Psychometrics, Model Evaluation, LLM Evaluation

Source: Open primary source (opens in a new tab)

6
Authors
9
Findings
20
Limitations
10
Evidence

Editorial summary

English

This Findings of EMNLP 2025 paper is a narrative survey and taxonomy proposal about personality in LLMs, not a reproducible systematic review or an empirical study. That distinction is central. Although the abstract says it systematically analyzes limitations and the introduction calls the work the first systematic account of multimodal personality modeling, the paper reports no bibliographic databases, search query, search date or time window, inclusion and exclusion criteria, screening process, number of retrieved or included studies, reviewers, extraction protocol, quality assessment, or risk-of-bias appraisal. It does not provide an included-study table. Comprehensive therefore describes topical ambition rather than demonstrated coverage, and systematic should not be read as a formal systematic-review design. The authors themselves acknowledge that recent or domain-specific contributions may be missing, that low-resource languages, cultural adaptation, and longitudinal user studies are outside scope, and that they perform no benchmarking, reimplementation, or quantitative comparison. The defensible contribution is a broad conceptual map. The taxonomy first organizes how personality is modeled. Early approaches include manual rules and lexicons, engineered features, n-grams, classical embeddings, and machine learning, followed by pretrained models such as BERT and GPT-2. In the LLM era it separates model-centric approaches, zero-shot, few-shot, and fine-tuning, from system-level approaches, RAG and agents. Zero-shot and few-shot control output through instructions or demonstrations. Fine-tuning changes model parameters and may offer greater stability, but requires data, compute, and safeguards against overfitting or catastrophic forgetting. RAG can add user or contextual information without retraining but depends on retrieval quality. Agents can maintain profiles, memory, and tools but make persona coherence more complex. This classification is useful for locating techniques, yet it groups distinct phenomena under personality: psychometric traits, induced styles, user profiles, personalization, fictional characters, role-playing, and agent behavior. The paper supplies no construct test for deciding when a conditioned output is a trait, a persona, or ordinary instruction following. A second dimension extends expression beyond text to vision, voice, and virtual reality. The argument is that prosody, facial expression, gesture, and embodiment can make personality more credible. However, the multimodal evidence is much thinner than the textual literature, and the VR subsection prominently relies on one 2025 study coauthored by survey authors. The limitations concede that most cited work remains text-centric. Embodying therefore marks a promising agenda, not a mature or demonstrably representative evidence base. The third dimension compares qualitative and quantitative evaluation. Human evaluation can provide contextual sensitivity but is expensive, subjective, difficult to reproduce, and sensitive to annotator demographics. LLM-as-judge reduces cost and scales judgment but imports evaluator bias and cultural limitations, can be inconsistent or hallucinate, and needs independent calibration. Among quantitative methods, psychometric questionnaires yield apparently standardized scores but were designed for humans. A model predicts tokens rather than introspecting a stable self; option order, wording, scale format, framing, and possible training exposure can change answers. LIWC is traceable and easy to apply, but rigid dictionaries lose context and semantics. Vectors and embeddings capture contextual relations and scale well, yet are less interpretable and can correlate with personality labels without psychometric grounding. This discussion correctly identifies important threats but empirically validates none of the method families. Table 1 assigns binary values to five methods for Traceable, Scalable, Prompt-Agnostic, and Context-Aware. The axes are neither operationalized nor measured. For example, whether human evaluation is prompt-agnostic or personality tests are context-unaware depends on the actual protocol. The table is an editorial heuristic rather than comparative evidence. The proposed directions follow logically from the diagnosis: develop prompt-invariant, context-aware, psychometrically grounded benchmarks; integrate modalities; scale personalization; and ensure conditioned behavior remains safe, coherent, and socially appropriate. The ethical discussion notes manipulation by emotionally adaptive personas, stereotypes imposed by trait taxonomies, impersonation of public or fictional figures, and lack of consent or transparency in longitudinal adaptation. These are pertinent risks, but they are not subjected to systematic evidence review or causal evaluation. The paper contributes no experiments, sample, executed models, new dataset, code, or quantitative synthesis. Claims about advances belong to cited primary studies and must not be represented as the survey's own findings. Nor does it establish that LLMs possess internal traits, that any technique produces stable personality, that human inventories are valid for machines, that the literature coverage is complete, or that there is consensus on the best evaluation method. It should be cited as a useful and critical narrative taxonomy: it supplies vocabulary connecting modeling, modalities, and evaluation, while the strength of each conclusion depends on the primary sources and completeness cannot be audited. The previous repository record also had a critical integrity defect: abstract_en_original was a generic invented paragraph absent from the publication. It has been replaced verbatim with the ACL abstract, including the source's development LLMs and consistent consistent errors because an original-text field must remain faithful to the source.

Español

Esta publicación de Findings of EMNLP 2025 es una revisión narrativa y una propuesta de taxonomía sobre personalidad en LLM, no una revisión sistemática reproducible ni un estudio empírico. Esa distinción es central. Aunque el abstract dice que analiza sistemáticamente las limitaciones y la introducción presenta el trabajo como el primer systematic account del modelado multimodal de personalidad, el artículo no informa bases bibliográficas, consulta de búsqueda, fecha o intervalo de búsqueda, criterios de inclusión y exclusión, proceso de cribado, número de trabajos recuperados o incluidos, revisores, protocolo de extracción, evaluación de calidad ni riesgo de sesgo. Tampoco ofrece una tabla de estudios incluidos. Por tanto, comprehensive describe la ambición temática, no una cobertura demostrada, y systematic no debe interpretarse como el diseño formal de una systematic review. Los propios autores reconocen que pueden faltar aportaciones recientes o de dominios específicos, que quedan fuera los idiomas de bajos recursos, la adaptación cultural y los estudios longitudinales, y que no realizan benchmarking, reimplementación ni comparación cuantitativa. La contribución defendible es un mapa conceptual amplio. La taxonomía organiza primero cómo se modela la personalidad. Los antecedentes incluyen reglas y léxicos manuales, rasgos construidos, n-gramas, embeddings clásicos y machine learning, seguidos por modelos preentrenados como BERT y GPT-2. En la etapa LLM distingue enfoques model-centric, zero-shot, few-shot y fine-tuning, de enfoques system-level, RAG y agentes. Zero-shot y few-shot controlan la salida mediante prompts o demostraciones; fine-tuning modifica parámetros y puede ofrecer mayor estabilidad, pero exige datos, cómputo y controles frente a overfitting o catastrophic forgetting. RAG puede introducir información de usuario o contexto sin reentrenar, aunque depende de la calidad de recuperación, y los agentes pueden mantener perfiles, memoria y herramientas, pero hacen más compleja la coherencia de la persona. Esta clasificación es útil para ubicar técnicas, pero reúne bajo personality fenómenos distintos: rasgos psicométricos, estilos inducidos, perfiles de usuario, personalización, personajes ficticios, role-play y conducta de agentes. El paper no fija una prueba que permita determinar cuándo una salida condicionada es un rasgo, una persona o simple instruction following. Una segunda dimensión amplía la expresión más allá del texto: visión, voz y realidad virtual. La tesis es que prosodia, expresión facial, gesto y embodiment pueden hacer más creíble una personalidad. Sin embargo, la evidencia multimodal es mucho más escasa que la textual y la sección de VR se apoya de forma destacada en un estudio de 2025 cofirmado por autoras de la revisión. Los autores admiten que la literatura citada sigue siendo principalmente text-centric. En consecuencia, Embodying señala una agenda prometedora, no una base madura o representativa de resultados. La tercera dimensión compara evaluación cualitativa y cuantitativa. Entre las cualitativas, la evaluación humana aporta sensibilidad contextual, pero es cara, subjetiva, difícil de reproducir y sensible a la composición demográfica. LLM-as-judge reduce coste y escala el juicio, pero hereda sesgos y limitaciones culturales del evaluador, puede ser inconsistente o alucinar y requiere calibración independiente. Entre las cuantitativas, los cuestionarios psicométricos producen puntuaciones aparentemente estandarizadas, pero fueron diseñados para humanos. El modelo predice tokens, no introspecciona un yo estable; orden de opciones, redacción, formato de escala, framing y posible exposición del material en entrenamiento pueden cambiar la respuesta. LIWC es trazable y fácil de usar, pero sus diccionarios rígidos pierden contexto y semántica. Los vectores y embeddings capturan relaciones contextuales y escalan bien, pero son menos interpretables y pueden correlacionar con etiquetas de personalidad sin estar fundamentados psicométricamente. Esta discusión identifica correctamente amenazas importantes, pero no valida empíricamente ninguna familia. Table 1 asigna a cinco métodos valores binarios para Traceable, Scalable, Prompt-Agnostic y Context-Aware. Los ejes no se operacionalizan ni se miden; por ejemplo, declarar human evaluation prompt-agnostic o personality tests no context-aware depende del protocolo concreto. La tabla es una heurística editorial, no evidencia comparativa. Los retos propuestos son coherentes con el diagnóstico: desarrollar benchmarks invariantes al prompt, sensibles al contexto y psicométricamente fundamentados; integrar modalidades; escalar la personalización; y asegurar que la conducta condicionada sea segura, coherente y socialmente apropiada. La discusión ética menciona manipulación por personas emocionalmente adaptativas, estereotipos impuestos por taxonomías de rasgos, suplantación de figuras públicas o personajes y falta de consentimiento o transparencia en adaptaciones longitudinales. Son riesgos pertinentes, pero tampoco se someten a una revisión sistemática o a evaluación causal. El artículo no aporta experimentos, muestra, modelos ejecutados, datos nuevos, código ni síntesis cuantitativa. Sus afirmaciones sobre avances proceden de los trabajos citados y no deben convertirse en resultados propios. Tampoco demuestra que los LLM posean rasgos internos, que una técnica produzca personalidad estable, que los inventarios humanos sean válidos para máquinas, que la literatura cubierta sea completa ni que exista consenso sobre la mejor evaluación. Debe citarse como una taxonomía narrativa útil y crítica: ofrece vocabulario para conectar modelado, modalidades y evaluación, pero la fuerza de cada conclusión depende de las fuentes primarias y la exhaustividad no puede auditarse. La ficha anterior contenía además un fallo grave de integridad: abstract_en_original era un párrafo genérico inventado que no figuraba en la publicación. Se ha sustituido literalmente por el abstract de ACL, conservando incluso los errores development LLMs y consistent consistent porque el campo original debe ser fiel a la fuente.

Research question

How can the literature on modeling, multimodal expression, and personality evaluation in LLMs be conceptually organized, and what limitations and future directions does that literature identify?

Method

Narrative literature review organized through a proprietary functional taxonomy: pre-LLM antecedents; model-centric and system-level techniques; textual, visual, vocal, and VR modalities; and human evaluation, LLM-as-judge, questionnaires, LIWC, and vectors. No systematic method of search, selection, extraction, or quality evaluation is reported, nor are experiments or reimplementations executed.

Sample: It does not apply an experimental sample. The publication cites selected literature, but does not report how many records it identified, examined, excluded, or included, nor does it define the documentary universe; therefore, coverage cannot be audited.

Findings

  • The current source is Findings of EMNLP 2025, Anthology ID 2025.findings-emnlp.506, DOI 10.18653/v1/2025.findings-emnlp.506, pages 9519-9532.
  • The 14 pages were rendered and visually inspected; SHA-256 364d2c49a493c1ab3a8e6b93a4660deaf2492d7b6fe8ac05441b54cc61bf12f6.
  • The record stored an invented generic abstract; it was replaced with the literal published abstract.
  • The taxonomy separates antecedents, model-centric LLM methods, system-level methods, multimodality, and evaluation.
  • Evaluation is divided into human evaluation, LLM-as-judge, psychometric questionnaires, LIWC, and vector-based methods.
  • The survey identifies prompt fragility, lack of standards, contextual dependence, limited multimodal integration, and weak psychometric grounding.
  • The paper itself acknowledges probable omissions, exclusion of several domains, and absence of benchmarking or quantitative comparison.
  • Future directions prioritize prompt-invariant and context-aware benchmarks, multimodality, scalable personalization, and safe alignment.
  • The ethical discussion covers manipulation, stereotypes, impersonation, consent, and transparency.

Limitations

  • No database or bibliographic search source is reported.
  • There is no query, keywords, search date, or time interval.
  • There are no explicit inclusion or exclusion criteria.
  • No screening, deduplication, reviewers, disagreements, or flow diagram are reported.
  • It is not stated how many studies were retrieved, excluded, or included.
  • There is no extraction protocol or study-outcome table.
  • Quality, risk of bias, or certainty of evidence is not evaluated.
  • The exhaustiveness of the cited corpus cannot be reproduced or audited.
  • The claims first survey and first systematic account are not supported by a systematic search of previous reviews.
  • Personality, persona, personalization, style, role-play, and agents are grouped without an established operational boundary.
  • Examples of primary works are narrated without weighing design, size, replication, or quality.
  • Table 1 uses binary categories that are neither defined nor measured.
  • There is no empirical benchmark among evaluation methods.
  • Multimodal and embodiment evidence is scarce relative to textual evidence.
  • The VR section highlights a study co-authored by the review's authors and does not discuss possible selection bias.
  • Low-resource languages, cultural adaptation, or longitudinal studies are not covered in depth.
  • Methods are not reimplemented, nor is a quantitative comparison produced.
  • No code, dataset, outputs, or extraction artifacts are published.
  • Coverage is exposed to omissions due to the rapid evolution of the field.
  • The abstract and conclusion use systematic with a stronger meaning than the described method.

What the study does not establish

  • It does not establish that the review is systematic in the methodological sense.
  • It does not demonstrate complete or representative coverage of the literature.
  • It does not establish that this is the first review of the field through a reproducible search.
  • It does not demonstrate that LLMs possess internal or stable psychological traits.
  • It does not empirically separate personality from persona, style, or instruction following.
  • It does not validate human questionnaires to measure LLMs.
  • It does not demonstrate that one evaluation method is superior to another.
  • It does not provide experimental results or its own quantitative effects.
  • It does not demonstrate maturity of multimodal or embodied personality.
  • It does not establish consensus on how to model or evaluate personality in AI.

Traceability

Scope: Full text

Version: Findings of EMNLP 2025, Anthology ID 2025.findings-emnlp.506, DOI 10.18653/v1/2025.findings-emnlp.506, pages 9519-9532, 14 pages

Consulted source: https://aclanthology.org/2025.findings-emnlp.506/

Review: Codex complete bilingual full-text fidelity pass using the published ACL 2025 paper, all-page visual inspection, taxonomy reconstruction, systematic-review-method audit, construct-boundary review, evaluation-method reconciliation, abstract provenance correction, limitation extraction, and reproducibility assessment; summaries written from the full paper rather than abstract keywords, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Instruments and metrics

  • Taxonomía funcional de modelado, multimodalidad y evaluación
  • Comparación conceptual de human evaluation, LLM-as-judge, personality tests, LIWC y métodos vectoriales
  • Discusión narrativa de retos, direcciones futuras y ética

Data used

  • No se construye corpus de revisión auditable ni se publica una tabla de estudios incluidos
  • No hay dataset, benchmark, outputs o extracción de evidencia nuevos

Evidence and location

  • Version, authors, venue, DOI, pages, and literal abstract: ACL Anthology record and published PDF page 9519 checked 15 July 2026
  • Scope, declared novelty, and taxonomy: Introduction and Figure 2, pages 9519-9521
  • Antecedents and model-centric/system-level techniques: Sections 2-3, pages 9520-9524
  • Vision, audio, and VR: Section 4, pages 9524-9525
  • Human evaluation, LLM-as-judge, questionnaires, LIWC, and vectors: Section 5 and Table 1, pages 9525-9527
  • Challenges and future directions: Section 6, page 9527
  • Ethical risks: Section 7, page 9527
  • Absence of benchmark and coverage limits: Limitations, page 9528
  • Audit of method and abstract integrity: reports/verification/article-189-survey-method-and-taxonomy-audit.json
  • Complete visual inspection: All 14 published PDF pages rendered and visually inspected on 15 July 2026